Shared Context Board

board-pattern-gist.md

Shared Context Board

A pattern for carrying context across work with AI agents through a shared markdown board, or "blackboard."

This is close in spirit to blackboard systems and stigmergic coordination: the shared surface carries enough state for separate actors to coordinate through the environment.

This is an idea file. It can be handed to an AI coding or research agent such as Codex, Claude Code, or another tool. The goal is to communicate the pattern clearly enough that you and your agent can adapt it to your own work.

Here, "agent" can simply mean a new chat or thread in a tool like Codex or Claude Code. It does not have to mean a custom autonomous agent or automation, though the same board can support those later.

The Core Idea

Most AI-assisted work lives in chat. Chat is good for conversation, but weak as the only working memory. Important context gets trapped in a thread: what matters now, what changed last time, which decision is still open, and where the real work lives.

This gets worse when you are managing a portfolio of work: products, hiring decisions, customer situations, operating projects, research bets, internal narratives, and team rituals. Each line of work has its own artifacts and its own half-finished reasoning. The board gives the human and the agent one shared place to see the live state without asking the chat thread to hold everything.

The board turns work state into an environmental signal. One chat can leave a trace; a later chat can read that trace and continue from there. The continuity lives in the workspace, not only in one conversation.

The board is a shared surface that sits between the human, the agent, and the files.

A good board answers, at a glance:

• which bodies of work are active
• what state each one is in
• what decision or judgment is needed next
• where the current source of truth or deliverable lives

The board is not the whole knowledge base. It is the map of active work. The actual documents, notes, source files, spreadsheets, drafts, and research artifacts stay in their proper folders.

The Grammar

A possible grammar:

Arena -> Focus -> Thread -> Breadcrumb -> Decision -> Deliverable

Element	Role
Arena	The folder or domain where the work lives
Focus	A current problem, outcome, or line of work
Thread	One active line of effort inside a focus
Breadcrumb	A short progress trace that makes the next session resumable
Decision	Durable reasoning worth preserving outside chat
Deliverable	The real output, linked from the board but stored in the right place

This kind of vocabulary gives the agent a stable way to route work, resume from prior state, update the right artifact, and avoid flattening everything into a generic to-do list.

A Minimal Setup

A setup can be as simple as a few plain files and domain folders that match the work you actually manage. This could look something like:

your-workspace/
  AGENTS.md                  # or CLAUDE.md, GEMINI.md, etc.
  board/
    Work_Board.md
    START_HERE_Board_Guide.md
  products/
    product-alpha/
    product-beta/
  projects/
    project-a/
    project-b/
  hiring/
    roles/
    candidates/
  decisions/
    architecture/
    budget/
  customers/
  research/
  operations/
  archive/

AGENTS.md can carry the thin always-on instruction: read the board at the start of meaningful work and update it when durable state changes.

Work_Board.md is the current shared surface. It has a summary table at the top and short focus details below.

START_HERE_Board_Guide.md can explain the pattern, modes, and update loop. This file works best when it stays short. If the guide becomes a manual, the system will decay.

The domain folders are where the actual work lives. A product leader might have one area for each product, a projects area for cross-functional programs, a decisions area for durable tradeoffs, a customer area for account context, and a research area for market or technical exploration. A personal workspace could include domain folders like personal-admin/, personal-website/, research/, financial-stewardship/, and so on.

Board Shape

A simple board shape could have two layers.

First, a summary table. The Details column can link to the longer focus section below:

Focus	Arena	Status	Mode	Next decision	Updated	Details
Example project	projects/example/	active	produce	decide what to finish first	2026-06-21	Details

Then, one short section per focus:

### Focus: Example project
<a id="focus-example-project"></a>
- Arena: `projects/example/`
- Status: `active`
- Mode: `explore` | `decide` | `produce` | `refine`
- Desired output:
- Source links:
- Deliverable location:
- Next decision needed:

#### Active threads
- [ ] Thread 1
- [ ] Thread 2

#### Latest breadcrumbs
- YYYY-MM-DD - mode - action taken
  learned:
  next:

The summary row and focus section should move together. If the summary says the next decision is one thing and the focus section says another, the board can stop being reliable.

When a focus is done, mark it complete or move it to an archive section.

Modes

Modes are a small hint about what kind of help is useful right now.

Mode	Good for	Agent tendency
explore	the scope or problem is unclear	gather, compare, map the space
decide	tradeoffs matter	structure options and reduce ambiguity
produce	the shape is known enough to act	create or update the artifact
refine	an artifact already exists	tighten, simplify, standardize

The mode gives the agent a cue for the kind of help that fits the work. An explore focus invites mapping and comparison. A produce focus invites forward motion on an artifact that already has enough shape.

Starting A Session

In my setup, the human can start a durable-work chat with an ordinary question or statement: "let's pick back up on research topic X," "add this new source to the question we were working through," or "where did we leave the thread on Y?" The agent first orients itself from the board:

1. It reads board/Work_Board.md.
2. It reads board/START_HERE_Board_Guide.md.
3. It matches the request to an existing active focus when possible.
4. It creates a new focus when the request starts a durable line of work.
5. It works in the real artifact, not in the board.
6. It adds a short breadcrumb when durable progress happens.
7. It updates the summary and focus detail together.

Tiny standalone questions can skip the board. The boundary I use is: if there is no likely future relevance, no artifact, no decision trail, or it is basically a stray Google search, it can stay off-board.

What The Agent Picks Up

When this is set up well, the human does not need to narrate the workflow. They can start with the kind of request they would naturally make: "pick back up on the research question about X," "fold this new note into the thread on Y," or "what did we learn last time about Z?"

The agent uses the board to orient itself:

1. Choose or create a focus.
2. Set the mode.
3. Work one thread.
4. Update the deliverable where it belongs.
5. Leave a breadcrumb.
6. Record a decision if something became durable.
7. Link the deliverable from the board.

Breadcrumbs work best when they are short. They are progress traces, not diary entries. They exist so a future agent can resume without asking the human to reconstruct the last session.

Example breadcrumb:

- 2026-06-21 - produce - drafted the public board-pattern gist
  learned: the portable version describes the grammar and session habit without depending on one tool
  next: review the gist against the longer field note, then publish if it still feels clear

Why This Works

The board turns work state into an environmental signal.

One chat leaves a trace: a focus, a mode, a link, a breadcrumb, a next decision. A later chat, whether in the same tool or a different one, reads the environment and acts from there. The continuity does not depend on one chat thread remembering everything.

Influences

This pattern borrows from a few places:

• Andrej Karpathy's LLM Wiki gist: an idea file that communicates a portable human-agent pattern rather than a fixed implementation.
• Ben Follington's claude-blackboard: a blackboard-style approach for preserving context across sessions and agents.
• Nichecraft's "Petri Dishes, Not Factories": the reminder that faster execution makes understanding and feedback more important.

Notes

Markdown is enough at the beginning. A database, plugin, search index, graph, or automation layer can come later if repeated use earns it.